Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A detection device for detecting a plurality of classes of object data, wherein the plurality of classes are merged step-by-step into a predetermined multilayer structure according to a similarity criterion and arranged at a bottom layer as the most finely classified classes, the detection device comprising: an input circuit configured to input data to be detected; and a processor configured to execute a cascade classifier comprising a plurality of level classifiers connected in series, the plurality of level classifiers configured to perform a classification respectively on the classes at respective layers in the predetermined multilayer structure according to a from-coarse-to-fine policy, and each of the level classifiers comprising strong classifiers, a number of which corresponds to a number of the classes being classified, wherein each of the strong classifiers comprises a group of weak classifiers, and each of the weak classifiers is configured to perform a weak classification on the data to be detected using a feature, each of the level classifiers comprises a shared features list, each feature in the shared features list is shared by one or more weak classifiers respectively belonging to different strong classifiers, and the weak classifiers using a same feature and belonging to different strong classifiers have parameter values different from one another, and the cascade classifier is configured to subject the data to be detected sequentially to judgment by each of the strong classifiers in each of the level classifiers so that when the data to be detected is judged by one of the strong classifiers of a level classifier as non-object data, the data to be detected is precluded from judgment by the other strong classifiers of the level classifier and sub-classes in the subsequent level classifiers, judge, for each of the level classifiers, whether there is a feature in the shared features list that is correlated only with the respective strong classifiers precluded from the judgment, and label the feature, when the cascade classifier judges that the feature is correlated only with the respective strong classifiers precluded from the judgment, as an invalid feature for which no feature value is further calculated.
A multiclass object detection device detects multiple object classes by first organizing them into a multi-layered structure based on similarity, with the most specific classes at the bottom. The device includes an input to receive data, and a cascade classifier processed by a processor. This classifier has multiple level classifiers in series, each classifying objects at different layers of the structure using a coarse-to-fine approach. Each level classifier has strong classifiers, each corresponding to a class. A strong classifier consists of weak classifiers, each using a feature to perform weak classification. Level classifiers share a feature list; weak classifiers using the same feature but in different strong classifiers have different parameters. The cascade classifier judges data sequentially; if a strong classifier rejects the data, subsequent strong classifiers and subclasses are skipped. Invalid features correlated only with rejected strong classifiers are labeled.
2. The detection device according to claim 1 , wherein each of the level classifiers is further configured to calculate feature values of respective valid features in the shared feature list for the data to be detected, and to query, for each of the strong classifiers in the level classifier, a feature values list obtained from calculation according to the features used for the strong classifier to thereby determine and sum up outputs of the respective weak classifiers of the strong classifier to obtain a final output of the strong classifier.
The object detection device as described previously also calculates feature values for valid shared features within each level classifier. For each strong classifier, it queries a feature value list based on the features used by that strong classifier. It then determines and sums the outputs of the strong classifier's weak classifiers to get a final output for that strong classifier, indicating whether the object is part of that class.
3. The detection device according to claim 1 , wherein the cascade classifier is configured to terminate the classification and judge the data to be detected as non-object data when the data to be detected is rejected by all of the strong classifiers in any one of the level classifiers.
The object detection device as described previously is configured to stop classification and classify the input data as non-object data when all the strong classifiers in any one of the level classifiers reject the data. This early termination prevents unnecessary computation.
4. The detection device according to claim 1 , wherein a last one of the level classifiers further comprises a judgment unit configured to judge the data to be detected as having an object-class-attribute corresponding to a certain strong classifier when the data to be detected is passed by the strong classifier, and to judge the data to be detected as having a corresponding multi-object-class-attribute when the data to be detected is passed by the plurality of strong classifiers of the last one of the level classifier.
In the object detection device as described previously, the last level classifier has a judgment unit. If the data passes a strong classifier, the judgment unit labels the data with the object class attribute corresponding to that strong classifier. If the data passes multiple strong classifiers of the last level classifier, the judgment unit labels the data as having corresponding multi-object class attributes, indicating that the object belongs to multiple classes.
5. The detection device according to claim 1 , which is for detecting a plurality of classes of a predetermined object in an input image or video, further comprising: a window traversal circuit configured to perform window-traversal on an image to be detected or on an image intercepted from an video to be detected, and the cascade classifier is configured to perform a classification on a window image acquired by the window traversal circuit and to record a position and a size of the window image and all the object-class-attributes thereof when the window image is judged as an object class.
The object detection device as described previously is designed to detect multiple classes of a specific object within an input image or video. A window traversal circuit scans the image or a video frame. The cascade classifier then classifies each window image. If a window is classified as an object class, the device records the window's position, size, and all its object class attributes, effectively locating and identifying objects within the image/video.
6. A detection method for detecting a plurality of classes of object data, wherein the plurality of classes are merged step-by-step into a predetermined multilayer structure according to a similarity criterion and arranged at a bottom layer as the most finely classified classes, the detection method comprising: inputting data to be detected; and performing a classification on the data to be detected with a cascade classifier comprising a plurality of level classifiers connected in series, the plurality of level classifiers configured to perform a classification respectively on the classes at respective layers in the predetermined multilayer structure according to a from-coarse-to-fine policy, and each of the level classifiers comprising strong classifiers, a number of which corresponds to a number of the classes being classified, wherein the performing a classification with the cascade classifier comprises subjecting the data to be detected sequentially to judgment by the respective strong classifiers in the respective level classifiers, each of the level classifiers comprises a shared features list, each feature in the shared features list is shared by one or more weak classifiers respectively belonging to different strong classifiers, and the weak classifiers using a same feature and belonging to different strong classifiers have parameter values different from one another, the subjecting the data to be detected sequentially to judgment by each of the strong classifiers in each of the level classifiers includes, when the data to be detected is judged by one of the strong classifiers of a level classifier as non-object data, the data to be detected is precluded from judgment by the other strong classifiers of the level classifier and sub-classes in the subsequent level classifiers, and the subjecting the data to be detected sequentially to judgment by the respective strong classifiers in each of the level classifiers further comprises: judging whether there is a feature in the shared features list of the level classifier, which is correlated only with the respective strong classifiers precluded from the judgment; and labeling the feature, when the feature is correlated only with the respective strong classifiers precluded from the judgment, as an invalid feature for which no feature value is further calculated.
A method for detecting multiple object classes involves arranging classes into a multi-layered structure based on similarity, with the most specific classes at the bottom. The method includes inputting data, then classifying it using a cascade classifier consisting of level classifiers in series. Level classifiers classify objects at different layers of the structure using a coarse-to-fine approach. Each level classifier contains strong classifiers, each corresponding to a class. Classification involves sequentially judging data by strong classifiers in each level classifier. Level classifiers have a shared feature list; weak classifiers using the same feature but in different strong classifiers have different parameters. If a strong classifier rejects the data, subsequent classifiers and subclasses are skipped. Invalid features only associated with rejected strong classifiers are marked as invalid, and their feature values are not calculated.
7. The detection method according to claim 6 , wherein the subjecting the data to be detected sequentially to judgment by the respective strong classifiers in each of the level classifiers comprises: for the input data to be detected, calculating feature values of respective valid features in the shared features list of the level classifier; and for each of the strong classifiers in the level classifier, querying a feature values list obtained from calculation according to the features used for the strong classifier to thereby determine and sum up outputs of the respective weak classifiers of the strong classifier to obtain a final output of the strong classifier.
In the multiclass object detection method as described previously, the data is sequentially judged by the strong classifiers in each level classifier. This process includes calculating feature values for valid shared features within each level classifier. For each strong classifier, a feature value list is queried, based on the features used by that strong classifier. The outputs of the strong classifier's weak classifiers are determined and summed to obtain a final output for that strong classifier, signifying whether or not the object belongs to that particular class.
8. The detection method according to claim 6 , wherein the performing a classification with the cascade classifier further comprises: terminating the classification and judging the data to be detected as non-object data when the data to be detected is rejected by all of the strong classifiers in any one of the level classifiers.
In the multiclass object detection method as described previously, the classification process is terminated, and the input data is labeled as "non-object data" when all of the strong classifiers in any of the level classifiers reject the data. This early termination step saves computational resources.
9. The detection method according to claim 6 , further comprising, after classification with a last one of the level classifier: judging the data to be detected as having an object-class-attribute corresponding to a certain strong classifier when the data to be detected is passed by the strong classifier; and judging the data to be detected as having a corresponding multi-object-class-attribute when the data to be detected is passed by the plurality of strong classifiers of the last one of the level classifier.
In the multiclass object detection method as described previously, after the final level classifier performs classification, a judgment is made: if the data passes a certain strong classifier, it's labeled with that classifier's object-class attribute. If data passes multiple strong classifiers of the last level classifier, it is then labeled as having a corresponding multi-object-class attribute.
10. The detection method according to claim 6 , which is used for detecting a plurality of classes of a predetermined object in an input image or video, further comprising: performing a window-traversal on an image to be detected or on an image intercepted from a video to be detected, and the performing a classification on the data to be detected with the cascade classifier comprising performing a classification on a window image acquired by the window-traversal with the cascade classifier and recording a position and a size of the window image and all the object-class-attributes thereof when the window image is judged as an object class.
The multiclass object detection method as described previously is used for detecting multiple classes of an object in an image or video. The method involves performing window traversal on an input image or video frame. The cascade classifier then classifies each window image obtained by this traversal, and if it determines the window is an object class, the method records the window's position, size, and all associated object class attributes.
11. The detection method according to claim 10 , further comprising merging windows in a local vicinity with the object-class-attributes generated by the window traversal component.
The object detection method as previously described, which uses window traversal to identify objects, includes an additional step of merging windows that are close to each other and have similar object-class-attributes. This merging step helps refine the detection results by combining multiple detections of the same object into a single, more accurate detection.
12. The detection method according to claim 11 , wherein the merging the windows in the local vicinity comprises: for the windows with a position adjacent the window center, a similar size scaling and the same object-class-attributes, calculating an average central position and an average window size of a cluster of object windows in the vicinity, and taking the number of merged windows as a confidence of the result of merging; merging the object attributes of the results of merging with an adjacent central position and a similar size after merging, wherein when there are a plurality of results of merging with different object attributes in the vicinity of a certain position in the image, a sum of confidence of the respective object attributes a calculated, one of the object attributes with the largest sum of confidences is taken as a final object attribute, and a sum of the sums of confidence of the respective object attributes is taken as a confidence of the final result of merging; when the confidence of the final result of merging is above or equal to a preset threshold of confidence, the final result of merging is accepted; and when the confidence of the final result is below the preset threshold of confidence, the final result of merging is discarded.
In the window merging process described previously, windows with adjacent centers, similar scaling, and the same object-class attributes are merged. An average center position and size are calculated for a cluster of object windows, and the number of merged windows becomes the confidence score. These merged objects are merged again with adjacent, similarly sized windows. If multiple merged results with different object attributes exist near a certain image position, the confidences for each object attribute are summed, and the attribute with the largest sum is selected as the final attribute. The confidence of the final merged result is compared against a preset threshold. Results below the threshold are discarded.
13. A non-transitory computer-readable medium including executable instructions that when executed by a processor cause the processor to execute a detection method for detecting a plurality of classes of object data, wherein the plurality of classes are merged step-by-step into a predetermined multilayer structure according to a similarity criterion and arranged at a bottom layer as the most finely classified classes, the detection method comprising: inputting data to be detected; and performing a classification on the data to be detected with a cascade classifier comprising a plurality of level classifiers connected in series, the plurality of level classifiers configured to perform a classification respectively on the classes at respective layers in the predetermined multilayer structure according to a from-coarse-to-fine policy, and each of the level classifiers comprising strong classifiers, a number of which corresponds to a number of the classes being classified, wherein the performing a classification with the cascade classifier comprises subjecting the data to be detected sequentially to judgment by the respective strong classifiers in the respective level classifiers, each of the level classifiers comprises a shared features list, each feature in the shared features list is shared by one or more weak classifiers respectively belonging to different strong classifiers, and the weak classifiers using a same feature and belonging to different strong classifiers have parameter values different from one another, the subjecting the data to be detected sequentially to judgment by each of the strong classifiers in each of the level classifiers includes, when the data to be detected is judged by one of the strong classifiers of a level classifier as non-object data, the data to be detected is precluded from judgment by the other strong classifiers of the level classifier and sub-classes in the subsequent level classifiers, and the subjecting the data to be detected sequentially to judgment by the respective strong classifiers in each of the level classifiers further comprises: judging whether there is a feature in the shared features list of the level classifier, which is correlated only with the respective strong classifiers precluded from the judgment; and labeling the feature, when the feature is correlated only with the respective strong classifiers precluded from the judgment, as an invalid feature for which no feature value is further calculated.
A non-transitory computer-readable medium stores instructions for performing a multiclass object detection method. The method involves arranging classes into a multi-layered structure based on similarity, inputting data, and classifying it using a cascade classifier consisting of level classifiers in series. Level classifiers classify objects at different layers using a coarse-to-fine approach, and have strong classifiers, each corresponding to a class. Classification involves sequentially judging data by strong classifiers in each level classifier. Level classifiers have a shared feature list; weak classifiers using the same feature but in different strong classifiers have different parameters. If a strong classifier rejects the data, subsequent classifiers and subclasses are skipped. Invalid features only associated with rejected strong classifiers are marked as invalid, and their feature values are not calculated.
Unknown
September 23, 2014
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.